<p>The escalation in vehicle ownership has led to frequent occurrences of target occlusion in vehicle multi-target tracking and challenges such as multi-scale variations and adverse lighting conditions further complicate vehicle tracking. While numerous tracking methods have been proposed to address these challenges, many prioritize enhancing Multiple Object Tracking Accuracy (MOTA) at the expense of real-time performance, leaving room for further performance enhancements. This study proposed a vehicle multi-target tracking approach, that aims to strike a balance between accuracy and real-time efficiency. The method is improved based on CenterTrack and tailored for intricate road environments. By replacing the DLA backbone network with an enhanced EfficientNetV2, the model can extract more precise target features. Moreover, appearance features are integrated to supplant the tracking branch of the baseline method, alongside a multi-level trajectory association strategy to mitigate false associations in complex road settings. To bolster object consistency across frames and enhance object discrimination, an attention-based appearance reconstruction module is introduced. The proposed method is trained and assessed on a standard multi-target tracking dataset, achieving a MOTA of 73.8% and IDF1 of 80.8%. These results underscore the applicability of the proposed vehicle multi-target tracking method in complex road scenarios.</p>